Hy-MT2-1.8B-JP-Manga-Finetune-v3-multilingual

This repo contains a version of the Hy-MT2-1.8B model that's been finetuned across multiple languages to preserve or improve the original Hy-MT2-1.8B model's language translation capabilities beyond just Japanese to English, targeting several (but not all) of the original supported languages by Hy-MT2-1.8B. Translation for Simplified Chinese, Vietnamese, Italian, Indonesian, Korean, French, German, Spanish, Portuguese, Russian, Ukrainian, Thai, Hindi, Farsi and Arabic is improved or preserved relative to the base model. Polish and Turkish regress slightly.

On Japanese to English, v3 is behind v2. An earlier version of this card called the two "basically equal"; later measurement showed that was wrong. On held-out manga dialogue — this model's actual domain — v3 is roughly 4 chrF++ behind v2: 900 bubbles neither finetune's teacher had ever translated, references generated fresh, paired bootstrap over 2,000 resamples, giving 42.07 → 38.21 on recogniser output and 45.96 → 41.72 on clean text. A separate held-out English manga probe put the gap at 5.7. The "basically equal" reading came from a broader English mix, where the gap is about 1 chrF++ and mostly quantization.

Use v2 if English is your only target, v3 for anything else. The v2 finetune had collapsed the base model's non-English output — pooled across 21 languages it went off-target 44.5% of the time against the base model's 8.6%, and for Simplified Chinese 99.7%, meaning it essentially never produced Chinese at all. v3 brings that to 7.4%, marginally below the base model itself. (Off-target figures are same-detector comparisons between checkpoints; the absolute values carry a large noise floor on short text and should not be read on their own.)

Note that for the non-English targets this is largely a matter of preserving the base model's capability, which the v2 finetune had degraded — not of adding new in-domain training data for those languages. Expect quality on those pairs to be in the neighbourhood of stock Hy-MT2-1.8B, with the domain characteristics described below layered on top.

Scope: this is a manga finetune

The training data is Japanese manga dialogue, and the model is built for translating manga — short, self-contained lines, translated one line at a time. This applies to every target language, not just English.

It will still translate other Japanese, but the finetune's biases follow its training domain, and they are worth knowing before deploying it on something else:

• Fragmentary input gets completed rather than translated. Manga lines are short and elliptical while the training targets are complete, natural sentences, so the model learned to produce a whole utterance from a partial one. Give it truncated input — ASR output, a cut-off caption, a mid-sentence chunk — and it will confidently finish the thought instead of translating only what it was given. Outside the intended domain this is the most common failure.
• Real-world proper nouns are weaker than in the base model. The names in the training data are fictional character names. For public figures, places and brands, stock Hy-MT2-1.8B is often more reliable.
• Technical, business and other specialist terminology is not covered. Nothing in the training data teaches it.
• Register leans conversational. Formal or expository source text comes out more casual than the source warrants.

If your input is continuous speech or long-form prose rather than discrete lines of dialogue, A/B it against stock Hy-MT2-1.8B before committing — this finetune does not dominate the base model everywhere.

Prompting

The finetune was trained with a short terminology block in the prompt: a few reference translations, written in the target language, placed ahead of the Japanese text to translate. The block is keyed to the target language, and the model does attend to it — ablating it on the earlier English-only finetune measurably lowered quality (about 1 chrF++, with some conspicuous individual failures). This is a real difference from stock Hy-MT2-1.8B, which largely ignores terminology injected into the prompt. If you have a glossary, putting it in the prompt is worthwhile here.

The instruction frame around that block was English during training; only the terminology block itself is written in the target language. Translating the entire prompt into the target language moves you further from the training distribution rather than closer to it.

Sampling

Recommended: temperature 0.15, top_k 20, top_p 0.6, repeat_penalty 1.05, min_p 0, with the sampler chain ordered penalties → top_k → top_p → temperature.

Higher temperatures measured worse. The cost is mostly variance rather than average quality, and that variance shows up as occasional invented content — so if you are seeing hallucination, lower the temperature before anything else. Note that llama.cpp and llama-server default min_p to 0.05; set it to 0 to match the conditions this model was evaluated under.